A hybrid CNN–transformer framework with self-supervised and federated learning for privacy-preserving cardiovascular disease diagnosis

Electrocardiography (ECG) is the most widely used diagnostic tool, but its effectiveness is limited by noise, inter-patient variability, and dependence on scarce expert-annotated data. This study proposes a hybrid diagnostic framework that integrates convolutional embeddings, Transformer encoders, self-supervised learning (SSL), and federated learning (FL) into a unified pipeline. Evaluations on PTB-XL, Chapman–Shaoxing, and MIT-BIH datasets demonstrate that the proposed model achieved a macro-F1 of 0.88 and AUC of 0.93, surpassing CNN (0.82 F1) and BiLSTM (0.81 F1) baselines. Under federated simulations with non-IID partitions, performance remained competitive (macro-F1 0.85, AUC 0.91), ensuring data privacy with only a 2–3% reduction from centralized training. Interpretability analyses confirmed attention focus on clinically relevant segments such as ST deviations and R–R irregularities. By unifying robustness, label efficiency, privacy, and clinical interpretability, this framework advances ECG-based CVD detection toward scalable, real-world deployment.